XZ-LI

AI Webnovel Pipeline — Data skill for Claude Code

Data community

Write a full-length Chinese webnovel with AI — layered memory for long-form consistency, 37 genre templates reverse-engineered from a 1,110-novel corpus.

How to install AI Webnovel Pipeline

This entry records only its repository, not the path inside it, so there is no exact command to give. Open XZ-LI/ai-webnovel-pipeline and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AI Webnovel Pipeline does

Write a full-length Chinese webnovel with AI — layered memory for long-form consistency, 37 genre templates reverse-engineered from a 1,110-novel corpus. Works in Claude Code / Cursor / Codex / Cline, or with no agent at all. 用 AI 写中文长篇网络小说：分层记忆保长篇一致性，37 个题材模版从 1,110 部作品语料反推而来，不用 agent 直接跑命令行也行

Alternatives in Data

LLM Redteam — LLM red-team corpus runner — fires categorized prompt-injection / jailbreak / system-prompt-leak / data-exfil 4.5k ★
HTML Video — Programmatic video for coding agents — HTML to video on your laptop 4.5k ★
Skill Content Pipeline — Extract patterns and anatomy from URLs — use to reverse-engineer content strategies from live pages 2.8k ★

README

ai-webnovel-pipeline

***English** · [简体中文](./README.zh-CN.md)*

Write a full-length Chinese webnovel with AI, and keep updating it.

You need no corpus of your own, and you do not need the plot worked out in advance. Once it is installed, say "I want to write a novel" in your terminal and it will walk you through pinning down the premise, hand you several opening options to choose from, set up the project, and then write chapter after chapter.

**Not tied to Claude Code**: it installs as a Claude Code skill, and Cursor / Codex CLI / Cline / Copilot / Windsurf / Aider all work too (see [below](#installing-into-other-ai-coding-tools)). It also works with no agent at all, straight from the command line — the pipeline is eleven Python scripts plus 37 genre templates, with zero agent dependency.

**The genre rules are reverse-engineered from a 1,110-novel corpus**, not hand-written prompts. See [where the rules come from](#where-the-rules-come-from).

What problem this solves

Writing one good chapter is easy. The hard part is that **chapter 200 still remembers what chapter 3 planted**. Six hundred thousand characters do not fit in any model's context window, so if you just keep saying "write the next chapter", by chapter twenty the protagonist's voice has drifted and the world rules contradict each other.

The core of this skill is a layered memory:

Layer	Stored as	What it holds	Refresh rate
Story bible	`bible.json`	World rules, character dossiers, overall arc, plus a set of lines that must never be crossed	Every 10 chapters
Foreshadowing ledger	`open_threads` in `bible.json`	What has been planted and not yet paid off, and in which chapter	Each chapter
Arc digests	`arc_digests.jsonl`	One paragraph per arc — how the distant past is remembered	When an arc completes
Arc plan	`arc_plan.txt`	What the current 30 chapters must deliver	When an arc completes
Summaries + recent prose	`summaries.jso